Google is giving Android developers a clearer way to see which AI models are actually useful when building Android apps, not just which ones look good on a generic coding leaderboard.
The company has updated Android Bench, its leaderboard for measuring how large language models handle real-world Android development tasks. The July release brings a new evaluation framework, refreshed scores, eight additional models, and a new way for developers to contribute their own testing scenarios.
For anyone using AI to help write, fix, or clean up Android code, this matters. A chatbot may be able to explain a function or suggest a quick snippet, but Android development brings its own set of wrinkles. Jetpack Compose migrations, wearable networking, platform API changes, and project-specific build issues can turn a simple prompt into a small debugging safari.
Android Bench is designed to measure how models perform in those Android-specific situations.
A New Framework for More Useful Results
Google says Android Bench now uses the Harbor framework as part of its July update. The benchmark previously used mini-swe-agent v1, a general-purpose benchmarking agent that was adapted for Android development tasks.
The move to Harbor gives Android Bench a more standardized foundation for running and sharing evaluations. In plain English, it should make the results easier to repeat, compare, and understand across different models and setups.
Google has re-run the benchmark across all models using the updated approach, so some scores have shifted. Older scores will still be available through the Android Bench archive for anyone who wants to compare past results.
For developers, the change means the leaderboard is getting a fresh measuring stick. The old one was not tossed into the junk drawer, but Google is clearly trying to keep pace with how quickly AI coding tools are changing.
Claude Fable 5 Leads the Updated Rankings
The July Android Bench release adds eight new models to the leaderboard: Claude Fable 5, Claude Sonnet 5, Claude Opus 4.8, GLM 5.2, Kimi K2.7 Code, MiniMax M3, Qwen 3.7 Plus, and Qwen 3.7 Max.
Claude Fable 5 currently sits at the top of the leaderboard with a score of 84.5. GPT 5.5 follows with a score of 80.2, and Claude Sonnet 5 lands in third with a score of 76.2.
Among open-weight models, GLM 5.2 leads with a score of 72.2, followed by Kimi K2.7 Code at 70.4.
Google says the leaderboard now includes performance and efficiency metrics, giving developers a better sense of how models compare beyond raw output quality. That could be helpful for teams trying to balance capability, cost, and speed without turning model selection into a spreadsheet dungeon.
Developers Can Help Shape the Benchmark
Google is opening Android Bench to more community input. Developers can now submit their own Android development tasks for review, giving the benchmark a better chance of reflecting the kinds of problems people run into during day-to-day work.
Developers can also run and share their own benchmark evaluations using Google’s dataset or custom tasks. Submitted tasks will be reviewed before they are considered for inclusion.
More details are available through the Android Bench website, GitHub repository, and Harbor Hub, where developers can review the dataset or submit evaluations.
